Long-term clothes-changing person re-identification

Abstract

Person Re-Identification(Person ReID) focuses on the searching and identifica- tion of pedestrians in cross-camera scenery. It can be regarded as a signif- icant complement to face recognition, providing sufficient features when face information cannot be obtained. Long Term Person Reid is a challenging sub- question that aims to match the same target for a long duration, and because it is a long-term problem, the individuals could be captured wearing different clothes. We assume that the face information is not accessible, because in the surveillance image, the faces are always very blurry or occluded. This disser- tation explores the performance of current representative methods on different datasets and sub-datasets, including more than 80 comparative experiments to compare the performance between using the human scale, gait features, seman- tically guided methods and ResNet50 baseline for comparison. In addition, dif- ferent datasets are also used for testing, including pure clothes changing datasets and hybrid datasets as well as different sub-datasets to test the performance of various methods. At the same time, based on the structure of ResNet50, this dissertation tries to add a mask structure to the building block, hoping to im- prove the model’s ability to perceive specific areas. In addition, for the semantic guidance method with better effect, this dissertation obtained a series of bench- mark results under different datasets. And it tested the influence of different data processing methods on the experimental results. Improvements were made and many meaningful and groundbreaking conclusions were drawn.